Study finds practice-session candidates accept 93% of AI-proposed code in coding interviews

·1 views

A study of 438 people practicing AI-assisted coding interviews on the nyrion.ai platform found that participants accepted 93% of the code the assistant proposed. Ten companies, including Canva, Shopify, Meta, Amazon and Google, have added an AI assistant to their coding interview rounds since June 2025. Candidates who received an AI proposal before first submit passed the stage's test suite 87% of the time, compared with 53% without one, though the author notes this is a correlation. When asked to explain their submitted code, only 9% of written answers were complete, while 79% were partial and 12% were off. The author stresses these are practice sessions, not real interviews, and that the figures reflect practice behavior rather than every company's format.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Five Beginner Web Development Mistakes That Cost One Developer Months

A developer reflects on early web development missteps that consumed months of effort, beginning with the belief that good developers must master many frameworks and code at high speed. They describe rushing from basic HTML and CSS straight into React without solid JavaScript fundamentals. Spreading attention across numerous tools led to tutorial overload and an inability to finish a complete project independently. The author also cites coding without planning, which produced tangled, hard-to-maintain projects, and skipping version control, which resulted in a lost folder and no public record of work. The piece concludes by urging beginners to build depth, plan ahead, use Git, and deploy their projects instead of leaving them on localhost.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Solo developer builds production SaaS in 40 minutes using AI software factory

A developer describes pastmonday.com, a small but complete SaaS app with authentication, a weekly calendar, Stripe billing, referral credits, file uploads, and an AI chat. The app was built by a 'software factory' of four reusable Claude Code skills that turn a one-page project brief into a shipped product in stages. Its backend runs almost entirely on Neon, which provides Postgres, Auth, and a Data API, with Object Storage and an AI Gateway in private preview that the developer accessed early. Stripe handles payments and the developer's own cluster hosts the containers. The developer says the skills are publicly available on GitHub and promotes the approach through a related video.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

New calculator estimates Apna Ghar loan eligibility from monthly salary

A developer has built a free online calculator for Pakistan's Apna Ghar Scheme that begins with a borrower's income rather than the desired loan amount, reflecting how banks assess applications. The tool supports salaried and self-employed modes, lets users set the share of income devoted to installments, and offers 5 to 20 year tenures. It also checks whether a property price falls within the scheme limit and estimates post-year-10 installments. The developer says a sample salary of PKR 80,000 yields about PKR 4.5 million, matching a real approval for a friend, but the tool is only an estimate, with banks verifying documents, credit history and current scheme terms.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

AI agent's startup file went stale three ways; its fix reports unknown instead of fresh

A long-running AI agent relies on a single startup file that names its identity, current phase, and documents to read first. That file went stale three times in one month, each time in a different way: a pointer to a replaced document, an outdated dated sentence, and a newer document it never listed. The agent wrote a small shell check covering these cases, which compares version paths, reads the date in the file against warning and red thresholds, and diffs directory contents against the listed entries. When the check cannot measure something, it reports unknown rather than fresh, and testing confirmed it flagged deliberately aged copies correctly. The author notes the approach resembles Nagios-style UNKNOWN exit codes and that similar tools exist, though this was tested only on one file.

0 comments Read more at DEV Community